Meta's Watermelon large language model remains in training as of late September 2026, with internal benchmarks reportedly reaching parity with OpenAI's GPT-5.5 on key evaluations according to Meta AI chief Alexandr Wang. The model uses roughly ten times the compute of the prior Muse Spark series released in April, signaling Meta's scaled-up investment in frontier capabilities amid competition from OpenAI, Anthropic, and Google. Recent reports indicate accelerated integration plans into the Muse agent platform following security stabilization, though no public release, weights, or Text Arena-style debut occurred at the September 23 Meta Connect event. Traders are watching for an October launch window and any verified third-party evaluations that could shift sentiment on competitive positioning.